Subgrid analysis of liquid jet atomization
Transcript of Subgrid analysis of liquid jet atomization
HAL Id: hal-00573786https://hal.archives-ouvertes.fr/hal-00573786
Submitted on 4 Mar 2011
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Subgrid analysis of liquid jet atomizationJeremy Chesnel, Julien Reveillon, François-Xavier Demoulin, Thibaut Ménard
To cite this version:Jeremy Chesnel, Julien Reveillon, François-Xavier Demoulin, Thibaut Ménard. Subgrid analysis ofliquid jet atomization. Atomization and Sprays, Begell House Inc., 2011, pp.41-67. �10.1615/Atom-izSpr.v21.i1.40�. �hal-00573786�
Subgrid analysis of liquid jet atomization
J. Chesnel a,∗, J. Reveillona, F.X. Demoulina, T. Menarda
aCNRS, CORIA UMR 6614, University of Rouen, Technople du Madrillet, BP 12, 76801
Saint-Etienne-du-Rouvray Cedex, France
Abstract
The objective of this paper is to study the feasibility of large eddy simulations of
a liquid fuel injection in combustion chambers. To do so, a priori analyses of direct
numerical simulations are carried out. A complete liquid jet atomization, from the
injector down to the end of the liquid core, is simulated thanks to the coupling of
both level-set and VOF formulations. To avoid the apparition of a subgrid term
in the right hand side of the continuity equation, the choice was made to consider
an incompressible formulation as far as the filtering operator is concerned. The
corresponding LES transport equations and various subgrid contributions are thus
presented. Results are first dedicated to the estimation of the various orders of
magnitude of these subgrid terms. In a second part, classical eddy viscosity scale
similarity models are tested against the prevalent ones. It appears that, contrary
to a Smagorinsky formulation, the scale similarity assumption provides a better
estimation of the subgrid terms. This result is found for all locations that have been
considered in the jet: at the injection level or in the atomized area. The major
drawback is the presence of a constant that needs to be estimated. Various values
are found depending on the filter size.
Keywords: atomization, subgrid modeling, Large Eddy Simulation, two-phase flow
∗Corresponding author. E-mail address: [email protected]
Preprint submitted to Atomization and Sprays November 29, 2010
Nomenclature
Roman letters
n Normal vector to the liquid/gas interface
uΓ Liquid/gas interface velocity vector
u Flow velocity vector
P Local Pressure
H Heaviside function
Greek letters
χ Liquid phase function
δΓ Dirac function centred on liquid/gas interface
κ Liquid/gas interface curvature
µ Local dynamic viscosity
Φls Level-set function
Φvof Volume Of Fluid (VOF)
ρ Local density
σ Surface tension coefficient
Operators
⋆ Spatial filtering operation
2
Introduction
Injection of liquid fuel in combustion chambers is the beginning of many physical
interactions involving complex phenomena, starting from the atomization process
down to the combustion and exhaust of burnt gases. Combustion regimes and flame
structures depend directly on the mixture fraction issued from the evaporation of the
liquid phase. Therefore, the atomization process is a key phenomenon controlling
the spray dispersion, evaporation and combustion in the chamber.
Development of economically efficient engines, while achieving low noise levels and
low pollutant emissions, is a challenge that needs modeling to be fulfilled. Experi-
ments and numerical simulations to study combustion in various application areas
have been carried out for several decades [Law, 2006] . Among them, modeling of
combustion devices has been at the center of numerous developments for forty years
owing to the importance of transportation vehicles [Baumgarten, 2006]. More re-
cently, the strong impact of the presence of a liquid phase prior to combustion has
been realized. Since the review paper of [Sirignano, 1983], modeling of evaporat-
ing sprays has made several advancements and several issues have been addressed,
although some stumbling blocks remain. By going ’upstream’ in the combustion
chamber, the first major physical phenomenon to be considered, analyzed and mod-
eled is the atomization process. However, because it involves a liquid/gas interface,
major difficulties appear from both experimental and numerical points of view and
the atomization process remains an open field to explore.
As far as large Weber and Reynolds numbers are concerned, as in many practical
situations, there is a lack of knowledge concerning the topology of a liquid jet and its
interactions with the surrounding turbulent flow leading to the spray formation. One
possible way to analyze the evolution of the liquid phase at the outlet of an injector
3
is to use direct numerical simulations (DNS), adapted to the presence of an inter-
face. When monophasic flows are considered, DNS solves the classical Navier-Stokes
equations, which consider the fluid as a continuum. Using a fine computational mesh
allows the macroscopic structures to be captured since all the considered distances
are considerably larger than molecular length and time scales. DNS was first intro-
duced 35 years ago by [Orszag and Patterson, 1972] followed by [Rogallo, 1981; Lee
et al., 1991] for the simulation of inert gaseous flows. It resolves all the characteristic
scales of a turbulent flow from the Kolmogorov ’dissipative’ length scale up to the
integral ’energy containing’ length scale. However, the apparition of an interface and
a strong variation of density jeopardizes the possibility of achieving a complete DNS.
This is especially true if fundamental physical phenomena, like evaporation or heat
transfers, are present at the interface level. Then, the computational cost of the DNS
of the whole flow, including both phases, would be skyrocketing unless some major
assumptions were made.
A first possibility is to adopt an interface tracking approach like the ’volume of
fluid’ (VOF) method developed by [Hirt and Nichols, 1981]. This approach is based
on the reconstruction of the gas/liquid interface from the time and space evolution of
the local volume fraction of liquid. This mass conservative procedure is complex as far
as the interface reconstruction is concerned. Another possibility is to use the level-set
(LS) procedure of [Osher and Sethian, 1988]. The procedure follows the motion of an
iso-surface of a specific scalar function that maintains algebraic distances. Generally,
an incompressible formulation is used but then, evaporation, heat transfer or even
combustion phenomena are difficult to account for. Nevertheless, these methods
are very promising as demonstrated recently by [Berlemont and Tanguy, 2005] who
simulated for the first time the complete atomization of a liquid jet.
DNS allows the surface tension forces of the liquid to be characterized, which
4
plays a major role in the atomization processes. These surface tension forces are
obtained from the estimation of the curvature of the liquid/gas interface, correctly
estimated by DNS, which captures the smallest scales of the flow. However, this
implies simulations with a limited domain size and low turbulent Reynolds numbers.
Furthermore, Reynolds Averaged Navier-Stokes (RANS) computations of atom-
ization can be carried out to determine the global impact of the sheared flow. RANS
methods are quick and necessary if we set out to model the whole combustion cham-
ber. However, numerous empirical laws are used to predict the spray granulometry,
the length of the liquid core and the spray velocity dispersion. This may lead to
strong errors as far as the atomization process is concerned, and it will affect all the
considered physical phenomena in the chamber. Progress has been made recently
by linking the primary breakup Eulerian modeling to a Lagrangian description of
the dispersed phase [Demoulin et al., 2007; Lebas et al., 2009; Vallet et al., 2001].
Yet, even if correct results are obtained in given cases, accounting for unstationnary
effects and large-scale instabilities remains a challenge.
To overcome these difficulties, the current trend in turbulence modeling is the
development of large eddy simulations (LES). LES is a promising technique between
industrial RANS solvers and academic DNS solvers. It solves the largest scales of the
flow while the impact of the smallest scales (or the subgrid scales) is modeled. Since
the development of subgrid dynamic models [Germano et al., 1991; Lilly, 1992], LES
of one-phase or two-phase dispersed flows is now widely used in the CFD community.
LES proved to be very efficient although it is more expensive than RANS computa-
tions. The idea of modeling turbulent flows and interfaces based on a LES procedure
is very recent and classic interface tracking methods have been used in LES solvers.
With the exception of [Yue et al., 2005], who used the Level-Set procedure, most of
the recent work uses the VOF method associated with a LES formulation to simulate
5
separated flows as waves [Lubin et al., 2006; Hieu et al., 2004; Christensen, 2006] and
bubbles [Deen et al., 2001; Liovic and Lakehal, 2007]. Although, in 2003, [De Villier
et al., 2004] have performed one of the first LES of primary diesel spray atomization,
the subject is rather unexplored. Moreover, major parts of LES studies that have
been carried out consider standard subgrid scale (SGS) models for the momentum
balance equation but not for the interface equation. Thus, the impact of the subgrid
flow properties on the droplet formation process is neglected and only the large-scale
instabilities participate in the spray formation. However, it should be noted that
[Chesnel et al., 2010] have proposed a first attempt at complete LES of atomization,
which takes into account the interface subgrid term.
Recently, [Labourasse et al., 2007] have performed the first analysis of the subgrid
term of the phase function transport equation. In their study, the authors conducted
a priori tests on DNS results to provide information on the order of magnitude of the
subgrid scale terms applied to bubble deformation. Their main conclusion was that
the most important subgrid scale term remains the inertia term, which can no longer
be modeled by a classical viscosity model close to the interface. In their subsequent
work, [Vincent et al., 2008] continued their analysis with a droplet of oil embedded
in water. Other a priori tests have been carried out in [Toutant et al., 2008] with
a strongly deformable bubble in a spatially decaying turbulence. This last reference
states clearly that a major difficulty arises when considering LES of flows containing
discontinuities like interfaces. Indeed, as detailed by [Sagaut and Germano, 2005],
subgrid models are generally based on the fact that gradients originate only from the
large-scale turbulence and not from the presence of an interface. One solution would
be, to set apart the interface impact on the turbulence model by using non-centered
filters. However, this procedure could be difficult to carry out from a technical point
of view. The solution adopted by [Labourasse et al., 2007; Toutant et al., 2008] was
6
to introduce the Interfaces and Subgrid Scales (ISS) as an equivalent to a one-phase
LES concept. In this framework, they used centered filters even at the interface
level. The subgrid terms issued from the discontinuity presence are accounted for in
models.
The objective of the presented research is to extend subgrid-scales a priori analy-
ses to complex two-phase flows thanks to the DNS of the full atomization of a liquid
jet. In this framework, the subgrid terms are evaluated in a complete atomization
process from the liquid injection down to droplets dispersion. Following the brief
presentation of the considered equations and the DNS of liquid jet atomization, we
test several closures for the so called subgrid terms. Results are provided with an
evaluation of the modeling of the momentum and phase function subgrid terms.
1. Modeling considerations
Since the first dynamic model developed by Germano [Germano et al., 1991], LES
is seen as the most promising tool concerning the modeling of turbulent flows. Much
research has been done for both incompressible and compressible flows. However,
as detailed in the introduction, LES of turbulent interfacial flows has rarely been
addressed. Thus, in the following, two-phase flow equations are presented and a
filtering operator is applied so that large eddy simulation formulation can be detailed.
Ultimately, mathematical formulations for the subgrid closures are presented.
It should be noted that indicial notations, via Einstein summation convention,
have been adopted instead of the more compact vector notations, since a non-
homogeneous sheared flow has been used in the following. For the sake of clarity, it
is important to show the various directions of the tensors.
7
1.1. The 1-fluid formulation
First and foremost, it is necessary to establish the governing equations of the
the two-phase flow. To do so, we start from the so called 1-fluid formulation in this
study. The main idea is to obtain a single set of governing equations, which allows
us to describe the whole simulation domain (liquid/gas in our case).
Considering two non-miscible fluids (indiced k = l and k = g for liquid and gas,
respectively), this formalism introduces a phase function, denoted χk, to discriminate
each phase and to allow for the mass and momentum jump conditions across the
liquid/gas interface. This variable, also called ’color function’, is defined by χk = 1
for phase k and χk = 0 otherwise. In this study, two phases are considered: liquid
and gas. Thus, for the sake of simplicity, χ represents the liquid phase and (1 − χ)
the gas phase. Thus, the local quantities are obtained using the phase function:
ρ = χρl + (1 − χ)ρg , (1)
µ = χµl + (1 − χ)µg , (2)
u = χul + (1 − χ)ug , (3)
P = χPl + (1 − χ)Pg , (4)
where ρ , µ, u and P are the local density, dynamic viscosity, velocity vector and
pressure, respectively.
Following the pioneering works of [Delhaye, 1974], [Drew, 1983] or [Kataoka,
1986], an advection equation is used for the phase function χ, in a classical manner:
∂χ
∂t+ uΓ
i
∂χ
∂xi= 0 . (5)
8
There is no mass exchange between the phases at the interface in our configu-
ration. Consequently, the interface velocity, uΓ, and the flow velocity, u, are both
equivalent.
Thus, considering the Navier-Stokes equations in each phase k, jump conditions
and relations 1 to 4, we can build the 1-fluid mass and momentum transport equa-
tions system. Especially for the considered case of incompressible, isothermal fluids
without evaporation phenomenon, this leads to:
• Mass∂ui
∂xi= 0 . (6)
• Momentum∂ρui
∂t+
∂ρuiuj
∂xj= −
∂P
∂xi+
∂Dij
∂xj− σκniδΓ , (7)
where:
Dij = µ
(∂ui
∂xj+
∂uj
∂xi
). (8)
The last term on the right hand side of equation 7 represents the surface tension
force. Since this force only acts at the liquid/gas interface, δΓ is the Dirac function
related to it. Thus, σ denotes the surface tension coefficient, κ and n represent the
curvature of the interface and the normal vector of the latter.
Finally, the total governing equations set is made up of equations 6 and 7 asso-
ciated with relation 5 for interface tracking. It should be noted that this physical
model is widely used in the DNS of turbulent two phase flows (e.g. [Menard et al.,
2007]). Consequently, many LES a priori studies, based on DNS databases, have
characterized the subgrid terms behavior, which appears when filtering these last
equations. We particularly mention the works of [Labourasse et al., 2007; Vincent
et al., 2008; Toutant et al., 2008] or more recently [Larocque et al., 2010].
9
1.2. Interface tracking
From the numerical point of view, resolution of equation 5 requires the develop-
ment of special procedures generally named interface tracking methods.
To track the interface, various numerical methods exist. A brief summary may
be found in [Lakehal et al., 2002] and reference therein. In the following, two of
the most used approaches are briefly described: the level-set and the volume of fluid
procedure. In fact, these two methods have been combined in order to obtain the
DNS database of this study.
1.2.1. Volume of fluid (VOF)
The VOF method uses a phase function Φvof (x, t) to characterize the volume
occupied by each phase in a cell. It is defined by the following equation:
Φvof (x, t) =Vl(x, t)
V, (9)
where V is the cell volume and Vl(x, t) =∫
Vχ(x, t)dV , the volume of liquid in the
cell. Therefore, cells with Φvof included between 0 and 1 contain the interface, the
pure liquid phase is designed by Φvof = 1 , whereas Φvof = 0 indicates a pure gas
phase. The evolution of the liquid volume fraction Φvof is characterized by an evolu-
tion equation similar to equation 5. The interface location is not directly known and
a reconstruction algorithm has to be applied to determine the local surface tension
force to be added in the momentum evolution equation. Even if the reconstruc-
tion procedure induces severe computational costs, the VOF method presents the
advantage of being mass conservative.
1.2.2. Level-set (LS)
The LS method uses a distance function Φls(x, t), which provides the algebraic
shortest distance of any point x to the interface. The sign of Φls depends on the
10
concerned phase ( Φls > 0 : liquid phase and Φls < 0 : gas phase). Thus, the interface
corresponds to Φls = 0. The liquid composition field may be retrieved through the
relation:
χl(x, t) = H(Φls (x, t)
), (10)
where H is the Heaviside function. As for the VOF procedure, the time and space
evolution of the distance function is governed by an equation similar to equation 5.
This method presents the great advantage of tracking the interface evolution and
thus, no reconstruction is necessary. However, mass conservation problems appear
frequently.
A promising solution is to associate both LS and VOF methods [Menard et al.,
2007]. By doing so, the VOF procedure ensures complete mass conservation of the
liquid while the LS method provides the interface position and suppresses the costly
reconstruction procedure, especially in 3D space. This is the procedure we adopted
for the DNS of a jet atomization presented in this study. Consequently, both usual
procedures have been combined.
1.2.3. Algebraic properties of the interface
Topological information about the interface is necessary to determine the con-
tribution of the surface tension forces on the momentum evolution of both phases.
Whichever phase function is used (Φvof or Φls), it is denoted Φ in the following.
In fact, for both formulations, similar expressions allow the definition of the vector
normal to the interface:
n =∇Φ
| ∇Φ |, (11)
and the interface curvature:
κ = −∇ · n . (12)
11
Then, Φ may be considered as the ’mollified color function’ as defined by [Brackbill
et al., 1992] or a ’smooth approximation’ as denoted by [Vincent et al., 2008].
In the following, surface tension forces have been modeled with the Continuum
Surface Force (CSF) method, proposed by [Brackbill et al., 1992], using the level-
set function, which offers better topological results than VOF (e.g. [Chesnel et al.,
2007]).
1.3. LES transport equations
1.3.1. Filtering operation
As pointed out in the introduction, a formal filtering operation should segregate
both phases in order to avoid spurious subgrid information due to the presence of the
interface among the turbulence models. However, in order to be able to develop and
use models in practical configurations with complex interfaces, it seems difficult to
apply filters that would have to be adapted along with the flow evolution. Thus, the
interfaces and subgrid scales (ISS) concept [Labourasse et al., 2007; Toutant et al.,
2008] seems more adapted: a classical filtering of the turbulent flow is carried out
and the interface presence is taken into account in the model definition.
Any field F (x, t) may be spatially filtered thanks to the relation:
F (x, t) = G ⋆ F (x, t) , (13)
which is a convolution operation over the computational domain volume V:
F (x, t) =
∫
V
G(∆,x − x′)F (x′, t)dx′ . (14)
The filter function G(∆,x) allows the definition of a cutoff length scale ∆. Several
shapes are possible. In our study, a classical top-hat filter has been used: G(∆,x) =
1/∆3
if | xi |≤ ∆/2 , ∀i ; otherwise, the filter function is null.
12
From a numerical point of view, the filtering operator used in this study is dis-
cretized following:
F (i, j, k) =1
(2N + 1)3
N∑
n=−N
N∑
m=−N
N∑
l=−N
F (i + n, j + m, k + l) , (15)
with:
2N + 1 =
(∆
∆
), (16)
where ∆ is the DNS mesh size. Thus, since ∆ is constant in time and space in
the DNS database, the filtering operator (eq. 15) commutes with time and spatial
derivatives [Sagaut, 2003].
Notes on Favre filtering.
As usual, when variable density flows are involved, a Favre filtering operation
may be defined
F (x, t) =ρF
ρ. (17)
Therefore, even if the density is constant in each phase, variations of density at
the interface level are characterized. However, as stated out by [Labourasse et al.,
2007; Toutant et al., 2008], when both phases are incompressible, the Favre filtering
operator leads to the disappearance of the divergence free expression (eq. 6).
Instead, a subgrid term appears on the right hand side of the continuity equation:
∂ui
∂xi=
ρl − ρg
ρ(δΓniui − δΓni ui) . (18)
13
This term is difficult to close and it could have a strong impact on the whole
simulation as noted by [Toutant et al., 2008]. Yet, the incompressibility hypothesis
remains identical if operator (.) is used:
∂ui
∂xi= 0 .
A simple way could be to neglect the right hand side of equation 18, as proposed
by [Liovic and Lakehal, 2007] on effective LES simulations. However, the preced-
ing authors mentioned the fact that this hypothesis should be used with precaution
and assume that this simplification can be made thanks to sufficient time and mesh
resolution, in order to capture the liquid/gas interface evolution. Moreover, concern-
ing a priori tests from DNS databases, it seems that this term also depends on the
flow configuration. For example, concerning the phase separation flow configuration,
[Larocque et al., 2010] has recently shown that this term cannot be neglected.
Thus, whenever possible, Favre filtering has not been used in this paper since no
simple solution may be suggested to close the right hand side of equation 18.
1.3.2. Filtered 1-fluid formulation
A filtering operator 14 is applied to equation 5:
∂χ
∂t+ ui
∂χ
∂xi= 0 , (19)
that can be rearranged in
∂χ
∂t+
∂ui χ
∂xi+ τχ = 0 . (20)
to extract the subgrid correlations between the interface and the velocity field
14
τχ = ui∂χ
∂xi
− ui∂χ
∂xi
. (21)
Physical properties of the flow in the vicinity of the interface depend on the
composition field χ. The filtered dynamic viscosity and density fields are defined by
µ = (1 − χ)µg + χµl , (22)
and
ρ = (1 − χ)ρg + χρl , (23)
respectively.
To determine the time and space evolution of both phases, Navier-Stokes equa-
tions are resolved. Each phase is incompressible and, as detailed previously, the
classic divergence free system may be written, once the filtering operator has been
applied:
∂ui
∂xi= 0 , (24)
∂ρ ui + τρui
∂t+
∂ρ ui uj + τρuiuj
∂xj= −
∂P
∂xi+
∂Dij + τDij
∂xj
− σκniδΓ︸ ︷︷ ︸τσi
, (25)
with Dij = µ(
∂ui
∂xj+
∂uj
∂xi
). The subgrid contributions issued from the filtering of
the Navier-Stokes equations are defined by the following expressions:
15
τρui= ρui − ρ ui , (26)
τρuiuj= ρuiuj − ρ ui uj , (27)
τDij= µ(
∂ui
∂xj+
∂uj
∂xi) − Dij , (28)
τσi= σκniδΓ . (29)
Using the name convention established by [Toutant et al., 2008], we shall use the
following denomination for the subgrid terms:
τχ : ”Interfacial term”
τρui: ”Temporal term”
τρuiuj: ”Convective term ”
τDij: ”Diffusive term”
τσi: ”Surface tension term”
All of these subgrid terms are due to two-phase flow consideration, except the
convective term, τρuiuj, which is present in standard single-phase LES equations. τρui
and τDijare due to the discontinuities of ρ and µ across the interface, respectively.
The two last terms, τχ and τσioriginate from subgrid interfacial deformations. This
can be better understood if we recall that ∂χk/∂xi = −nki δΓ in the 1-fluid formula-
tion.
On the one hand, previous a priori evaluations of these various terms suggest
that the convective term, τρuiuj, must systematically be taken into account through
a model. On the other hand, the behavior of the specific two-phase flow subgrid
terms strongly depends on flow configuration (see [Labourasse et al., 2007; Vincent
et al., 2008; Toutant et al., 2008; Larocque et al., 2010]).
16
2. DNS of the atomization of a liquid jet
Using the numerical procedure presented in [Tanguy et al., 2007], the DNS of the
complete atomization of a liquid jet has been carried out. This geometry presents
several advantages. First, it is very close to experimental and industrial consider-
ations where sheared flows are common. Indeed, jet injection is one of the main
ways to atomize a liquid into a spray of droplets and, as stated in the introduction,
there is a lack of information concerning the modeling of the area close to the injec-
tor. Therefore, some useful information concerning atomization modeling could be
derived from this configuration. Analyses are carried out starting from the undis-
turbed liquid sheet at the outlet of the injector down to a dense spray beyond the
disappearance of the liquid core. A wide range of various interface topologies is en-
countered. Thanks to the DNS, it is thus possible to study them and to evaluate the
capability of the models to capture their properties.
2.1. Configuration
The basis of Level Set methods has been proposed by [Osher and Sethian, 1988].
The interface is described with the zero level surface of a continuous function defined
by the signed distance to the interface. To ensure that the function remains the
signed distance to the interface, a re-distancing algorithm is applied. However, it is
well known that its numerical computation can generate mass loss in under-resolved
regions. To describe the interface discontinuities, a Ghost Fluid Method (GFM), as
derived by [Fedkiw et al., 1999], has been carried out to capture jump conditions
at the interface. The GFM not only avoids the introduction of a fictitious interface
thickness, but it is also suitable to provide a more accurate discretization of dis-
continuous terms, reducing spurious currents and improving the resolution on the
pressure jump condition ([Kang et al., 2000; Berlemont and Tanguy, 2005]). In the
17
primary break-up of a jet, numerous topological changes occur: interface pinching or
merging, droplet coalescence or secondary break-up. The numerical method should
describe the interface motion precisely and handle jump conditions at the interface
without artificial smoothing. Moreover, the method should be mass conservative.
In the computations carried out in this study, interface tracking is performed by a
Level Set method. GFM is used to accurately capture sharp discontinuities, and a
coupling between Level Set and VOF methods is made to ensure mass conservation
([Sussman and Puckett, 2000; van der Pijl et al., 2005]). A projection method is
used to solve incompressible Navier-Stokes equations, which are coupled to equation
5 for the level set function. A detailed description of the numerical methods may be
found in [Menard et al., 2007].
The main properties of the configuration are the following: the size of the do-
main is (2.4 mm, 0, 3 mm, 0, 3 mm), where the first dimension is the streamwise
direction and the other two, the spanwise directions (fig. 1). At the injection level,
the jet diameter D0 is equal to 0.1 mm, while the liquid jet Reynolds number is
equal to Re = 4659. Turbulence fluctuations are prescribed by using the method
developed by [Klein et al., 2003] with a characteristic correlation scale equal to a
tenth of the diameter of the injector. A summary of the physical parameters, for this
configuration, can be found in table 1.
Various stages of the injection process have been plotted in figure 2. Liquid
surface instabilities close to the injector are visible. Their deformation leads to the
formation of ligaments and droplets of various sizes. At the end of the domain,
the liquid core has almost disappeared and a dense spray of droplets leaves the
computational domain.
Concerning the smallest droplets, it is assumed that no breakup occurs below a
local gaseous Weber number value of 10. This critical Weber number, defined as
18
WeGc = ρgU2maxDgc/σ, leads to a minimum droplet diameter equal to Dgc = 2.4 µm.
Thus, a 2048 × 256 × 256 Cartesian grid has been used with regularly spaced nodes
(∆ = 1.17 µm).
In the following, a priori subgrid analyses are carried out to evaluate the capabil-
ity of LES formulations to capture the main characteristic of such an atomizing jet.
Three analysis areas have been defined: the injection area at a streamwise position
starting from 0 D0 to 4 D0 (zone 1), the central area positioned from 8 D0 to 12 D0
(zone 2) and eventually the atomized area positioned from 20 D0 to 24 D0 (zone 3).
These test zones are also visible in figure 1.
2.2. Phase function filtering
As defined in equation 14, a ’top-hat’ filtering operator is applied to the various
fields of the direct numerical simulation. To begin with, the composition field χ,
which characterizes the liquid phase (eq. 5), has been filtered. Two filter sizes have
been considered: a small filter size (SFS) ∆5 = 5∆ and a large filter size (LFS)
∆13 = 13∆, where ∆ is the uniform discretization step of the DNS Cartesian grid.
When scrutinizing figure 2, several questions arise concerning the capability of
LES models to reproduce such a system. Indeed, the objective of atomization mod-
eling is to provide accurate characteristics of the spray downstream the injector;
mainly, the droplet size and velocity distributions.
[De Villier et al., 2004; Buonfiglioli and Mendonca, 2005; Bianchi et al., 2007]
have conducted the first LES attempts of atomizing a liquid/gaz jet by using classical
closure for the convective subgrid term in the momentum equation and the interface
tracking method developed for DNS. In the following we will call this procedure
’Interface Coarse DNS’ (ICD).
However, actual use of LES for interfacial flows neglects the impact of subgrid
19
phenomena on the phase function and, though mass conservation is well respected,
the smallest droplets visible in figure 2 cannot be directly characterized.
Figure 3 presents a cut of the atomizing jet along the streamwise direction. It
shows the composition field χ in three cases: DNS, and then SFS and LFS filtering.
The top image presents the DNS composition field with the presence of primary liquid
instabilities and the formation of droplets and ligaments as described for figure 2. As
soon as a filter is applied, the neat border between liquid and gas phases disappears
and the smallest liquid structures vanish to be replaced by a smooth phase function
field that could be tracked on a mesh with much less grid points. We will focus on the
white lines in figure 3, representing the interface position that would be defined by
a large-scale computation. The key problem of large eddy simulations of interfacial
flows lies in this picture; in fact, the interface is represented only at a large-scale level.
Therefore, due to grid implicit filtering, the phase function frontier (white line) is no
longer able to capture the complete liquid evolution. This point is clearly exposed in
figure 4, which represents several flow density cuts perpendicular to the streamwise
direction. The first line of figures 4-(a-b-c) represents the flow density ρ, obtained
from the DNS computation for the three representative positions: injection, central
and atomized areas. In the framework of this analysis, this field may be considered
as an exact representation of the researched solution. The second line of figures 4-
(d-e-f) denotes the application of equation 23 defining the local density thanks to a
smoothed composition field. A lot of information is preserved and the presence of
ligaments and droplets at the subgrid level is still visible, even if the subgrid liquid
geometry is lost.
The limitation of using a composition field with an ICD formulation in a LES
computation is clearly visible in figures 4-(g-h-i). In these figures, the density field
is reconstructed using the sharp interface function related to larger scale, ε :
20
ρ(ε) = (1 − ε)ρg + ερl , (30)
For our purpose, ε is reconstructed thanks to the level 0.5 of χ :
ε = H(1
2− χ) =
1 if χ > 1/2
0 otherwise .(31)
It should be noted that choosing the χ = 0.5 level is an arbitrary assumption. In
ICD-LES simulations, mass conservation is guaranteed by the resolution of an ad-
vection equation for ε similar to equation 5 related to the interface tracking methods
developed for DNS.
In ICD formulation, only information about the main liquid core is known and,
because of mass conservation, inclusions smaller than the LES grid could not be
created and large liquid volumes are not ruptured. These conclusions emerge when
scrutinizing the shown ρ(ε) and ICD results in the field of liquid/gas atomization
[De Villier et al., 2004; Buonfiglioli and Mendonca, 2005; Bianchi et al., 2007]. In a
complete LES framework, the evolution of the interface cannot be treated without
taking into account the presence of the liquid phase at the subgrid level.
Clearly, two solutions can be envisioned. A first possibility, which has been
adopted in this study, is to work with a smoothed composition field, fig. 4-(d-e-f).
A second solution would be to use a classic composition field defining a large-scale
interface, as presented in figures 4-(g-h-i). This composition field could be charac-
terized thanks to a conventional VOF or LS phase function. The subgrid properties
of the liquid could be represented thanks to a new Eulerian variable defining the
subgrid liquid density. Owing to information provided by DNS, this quantity has
been plotted in figures 4-(j-k-l). However, an evolution equation would be necessary
and exchange terms with the large-scale composition field have to be determined.
21
Therefore, by associating fields 4-(g-h-i) and fields 4-(j-k-l) complete large-scale level
information can be obtained. In this paper, for the sake of simplicity, the use of a
smoothed composition field χ has been adopted. But the second solution could be
an interesting approach to explore, especially if industrial solvers are concerned.
Three main discussions arise concerning the LES of jet atomization. First of all,
we could question whether actual LES models are able to capture the correct time
and space evolutions of the filtered phase function field, such as the ones obtained
in figures 3 and 4. The objective of this paper is to provide a preliminary answer to
this question thanks to an a priori analysis. However, confirmation will be necessary
based on effective a posteriori computations. This will be the purpose of future
research. A second fundamental question appears when considering the liquid phase.
Indeed, one of the main objectives of atomization modeling is to provide information
about droplet size and velocity distributions. Therefore, it is necessary to reconstruct
this information from the LES field information. A possible method is to use a
’defiltering’ (also termed deconvolution) procedure [Sagaut, 2003], or to adopt a
statistical description of the properties of the subgrid liquid phase using data provided
by the filtered phase function. A last stumbling block concerns the destabilization of
the jet. In fact, surface instabilities appearing after injection drive the destabilization
of the jet. This leads to various atomization regimes depending on the injection
velocity, along with the liquid viscosity and density as well as the surface tension,
through the Reynolds and the Weber numbers [Reitz, 1978; Faeth, 1991]. As a
result, either LES should be refined to a DNS near the injector to capture the liquid
destabilization correctly, or specific subgrid models have to be developed to ensure
a correct jet destabilization.
22
2.3. A priori analysis of the subgrid terms
The objective of the following section is to evaluate the order of magnitude of
the subgrid terms present in the LES momentum and phase function equations. A
similar analysis has been conducted by [Labourasse et al., 2007] in the framework
of a bubble interacting with counter-rotating vortices and by [Vincent et al., 2008]
or [Larocque et al., 2010] for a water/oil inversion problem. An extension of their a
priori analysis is thus applied to the fully atomized jet in this study.
Thanks to the cylindrical symmetry, it is possible to carry out a radius depen-
dant analysis where data are averaged along the streamwise direction over a short
distance lx considered as homogeneous. The following averaging process has thus
been employed to analyze any field A(x, y, z) defined on the computational grid:
〈A〉 (r0, x0) =
∫ r0+lrr0
∫ x0+lx/2
x0−lx/2Ardrdx
∫ r0+lrr0
∫ x0+lx/2
x0−lx/2rdrdx
, (32)
where lx is the streamwise analysis length fixed to 4 D0, lr corresponds to the grid
step along the radial direction and the radius r =√
y2 + z2. A similar averaging
operator has been used to define interfacial statistics:
〈A〉Γ (r0, x0) =
∫ r0+lrr0
∫ x0+lx/2
x0−lx/2AδΓrdrdz
∫ r+lrr
∫ x0+lx/2
x0−xz/2δΓrdrdz
, (33)
To begin with, the mean turbulent kinetic energy 〈k〉 and the corresponding dis-
sipation rate 〈ε〉 have been plotted in figure 5. These data have been extracted from
the DNS fields; no filter has been applied yet. Close to the injector, turbulent kinetic
energy (fig. 5-a ) presents a usual profile with a peak appearing in the area where the
shear is at maximum. In our configuration, this corresponds to the interface position.
Then, while the jet is destabilized, turbulent structures develop and the global level
23
of kinetic energy decreases. The central and atomized area are thus the location of
turbulent energy production. In this part of the domain, it is possible to observe a
homogenization of the turbulence along the radial direction. Similarly to the kinetic
energy, the dissipation rate (fig. 5-b) presents a single peak in the injection area.
However, the dissipation level decreases along the streamwise direction while its pro-
file widens because of the jet destabilization and atomization. The final section of
our computational box corresponds to the end of the main liquid core atomization.
The domain is therefore too short to observe the decay of energy that will appear.
Logically, the corresponding level of dissipation remains of low energy production in
this area.
More details may be seen in figure 6, which represents the normalized turbulent
kinetic energy spectrum close to the injector, in the central area and in the atomized
area. Close to the injector, the turbulence is not yet fully developed but, very quickly,
an inertial range appears. Its slope respects the usual k−5/3 of the Kolmogorov theory.
When the filtering process is applied, it is possible to set apart the subgrid kinetic
energy from the resolved one, noted k. The proportion of resolved kinetic energy
compared to the total energy extracted from the DNS is plotted in figure 7 for the
two filter sizes SFS (∆5 = 5∆) and LFS (∆13 = 13∆), previously selected. In figure 6,
the SFS filter position corresponds to k/kmax = 0.2 and the LFS filter corresponds
to k/kmax = 0.076.
The left plot in figure 7-(a) shows that the smallest proportion of resolved energy
may be observed close to the injector in the liquid phase (r/r0 < 1). The gas
phase is still quiescent and a null energy may be measured (fig. 5-(a)). In the liquid
phase, even if the global energy remains weak compared to the peak observed at
the interface, turbulent motion is important because it participates in the surface
destabilization and atomization. If the SFS filter is applied, this ratio of large-scale
24
energy starts from 80 % in the liquid close to the injector and goes up to 90 % far
from the jet. This confirms that SFS filtering is useful for a progressive analysis
of subgrid filter scales but useless for an effective LES computation. Most of the
energy is resolved and the computational cost would be almost similar to a DNS
computation. However, as appears in figure 7-(a) if LFS is applied, only 20 % of
the energy is resolved at the center of the liquid jet outside the injector because
the largest turbulent liquid structures in the injector remain small compared to the
considered filter size. After injection, the jet expands and the size of the turbulent
structures increases in the liquid area (r/r0 < 1). Thus, resolved energy first increases
from 20 % to 50 % in the LFS profile in figure 7-(a). Then, because of the presence
of the interface (r/r0 = 1), a sudden drop down to 20 % of resolved energy may be
observed again. After this drop, resolved energy rises abruptly up to almost 100 %,
which corresponds to the quiescent atmosphere area. Far from the jet, it is possible
to observe (fig. 7-(b) ) that the flow becomes homogeneous along the radial direction
and the resolved energy reaches a constant 60 % proportion. A first remark concerns
the constant filter size that has been considered. Subgrid models have to reproduce
a wide range of unresolved energy: between 80 % in the liquid phase outside the
injector down to 40 % after atomization. This wide range implies a double constraint
on the filter size: It must be selected within the inertial range and it must be able
to represent subgrid energy in the whole calculation domain. Note that a similar
discrepancy between injection and far field exist also in monophasic jets and it has
been proven that LES is able to capture them [Wang et al., 2008]. However, we have
to keep in mind that, in our case, initial liquid perturbations contribute to the jet
atomization and the final size distribution of the droplets. The resolved energy drop
in the vicinity of the interface shows that small-scale structures are prevalent in this
area.
25
The complete subgrid budget of the momentum equation is considered in fig-
ure 8. In this figure, the four temporal 〈∂τρui/∂t〉, convective
⟨∂τρuiuj
/∂xj
⟩, diffusive
⟨∂τDij
/∂xj
⟩and interfacial terms 〈τσi
〉Γ defined in equation 25 have been plotted for
positions close to the injector (zone 1) and far from it (zone 3). Note that we chose
to work directly with the subgrid term τσi= σκniδΓ, since a resolved interface is not
clearly defined.
Initially, as for the turbulent kinetic energy evolution, the flow is strongly non-
homogeneous in the radial direction close to the injection. However, as soon as
atomization occurs, profiles along the streamwise direction become uniform (fig. 8-
(b) and (d)). Wherever we are along the streamwise direction, the hierarchy between
the various subgrid terms remains similar. At first, temporal and convective terms are
prevalent. The diffusion term is much smaller, with a difference of several orders of
magnitude. It appears that the interfacial term is very small compared to the others,
even the diffusion term. These remarkable properties have been observed previously
by [Labourasse et al., 2007] in the framework of the deformation of a bubble. In
their study, [Labourasse et al., 2007] suggested that the exceptionally high level of
the temporal term may be due to the low order of the time filtering derivative. They
were in fact using a first order Euler scheme. To check this hypothesis, it has been
decided in this study to use a fourth order temporal scheme. However, it appears
that similar conclusions may be drawn and the high magnitude of the temporal
term is not the consequence of the numerical scheme. In the work of [Labourasse
et al., 2007] the increase of the order of magnitude of the temporal scheme is directly
correlated to the deformation of the bubble: the more it is distorted, the higher the
temporal term. In our configuration, the liquid interface is much more deformed
than the considered bubble and thus, the temporal term remains strong wherever
we are in the streamwise direction. In fact, the more the interface is deformed, the
26
more it is possible to observe locally a brutal evolution of the flow density at the
subgrid level. Additionally, because of the high ratio of density between the liquid
and the gas phase, even a slight modification of the interface position may lead to
a high value of the temporal term, which represents the subgrid evolution of the
correlations between velocity and density.
Examples of convective terms close to the injector have been plotted in figure 9.
Figures 9-(a-b) represent a streamwise-spanwise term (τu1u2) and figures 9-(c-d) a
spanwise-spanwise term (τu2u2). Subgrid convective terms appear mainly around the
smoothed interface (represented by the thick isoline), corresponding also to the area
where the gradients reach their highest level. Instantaneous fields of the tensor have
been plotted to demonstrate that the interface presence has a remarkable effect on
the subgrid tensor. This is due to the appearance of small structures close to the
interface.
Resolved 〈u · ∇χ〉 and subgrid convective terms 〈τχ〉, of equation 20, have been
plotted in figure 10 for two positions: close to the injector and in the atomized
area. It appears that the interface convection is mainly controlled by large-scale
convection phenomena. In fact, the resolved part is often prevalent. However, the
subgrid contributes up to 10% of the resolved contribution, and is thus not negligible
for all the tested zones and filter sizes.
Several authors in previous works, [De Villier et al., 2004; Buonfiglioli and Men-
donca, 2005; Bianchi et al., 2007], have chosen to neglect the subgrid scales impact on
the interface propagation in liquid/gas atomization flow simulations. Consequently,
these authors highlight a strong impact of the mesh resolution on the jet destabiliza-
tion process as well as on the obtained droplet size distribution.
Close to the injector, the convective subgrid scale contributions could initiate
the initial interface destabilization process as noted by [Herrmann and Gorokhovski,
27
2008] even on large Weber number flows. As shown in figure 10-(a), for example,
there is a sudden growth of the subgrid term at the interface. It appears that the
interface subgrid term τχ may necessitate a model, at least in the atomization area.
Starting from the conclusions drawn in the previous section, we propose to test
some usual closure for the predominant subgrid terms in the following section.
3. Subgrid modeling
In order to carry out large eddy simulations of two-phase flows, subgrid terms,
defined in the preceding section, have to be modeled. The aim of this section is
to conduct some tests to know which closure should be used for the large eddy
simulation of two-phase flows.
We want to evaluate the subgrid term contributions as they appear in the evolu-
tion equations of the phase function (eq. 20) and the momentum (eq. 25). Following
the estimation of the subgrid terms importance in the previous section, we then fo-
cused on the closure of the prevalent terms of the momentum and phase function
equations:
• Temporal term contribution, Ci :
Ci =∂(ρuiuj − ρ ui uj)
∂xj
. (34)
• Convective term contribution, Ti :
Ti =∂(ρui − ρ ui)
∂t. (35)
• Convective term of the phase function equation contribution, Π :
Π = ui∂χ
∂xi− ui
∂χ
∂xi. (36)
28
It should be noted, that results are presented in dimensionless form, using the factors
presented below:
C∗
i = γCi ; T ∗
i = γTi ; Π∗ = ξΠ , (37)
where γ = D0/ρlU2d , ξ = D0/Ud and ξ = D0/Ud. For the sake of clarity, the
dimensionless notations (∗) are dropped. Thus, results and graphics presented in the
following are systematically non dimensional.
At first, the exact contribution is plotted against the modeled one in dispersion
diagrams for each tested model. From a practical point of view it is not possible
to plot all the points in each zone. Two data vectors were then used De (exact)
and Dm (model), containing n statistical sample points, determined in a random
manner. The kth component of these vectors contains exact and modeled subgrid
contributions taken at the same space point X, and time t..
3.1. Preliminary results
In the framework of one-phase turbulent flows, classic eddy viscosity models are
generally suggested to close the convective term. However, it appears that they
are ill-adapted to model subgrid scales of two-phase flows. In their pioneering work,
[Labourasse et al., 2007] reach the same conclusions after having tested the Smagorin-
sky model[Smagorinsky, 1963], the Wale model [Nicoud and Ducros, 1999] and mixed
scale models [Sagaut et al., 1999]. In addition, a dynamic Smagorinsky model was
tested [Germano et al., 1991] in the framework of our jet atomization configuration
and results were far from being satisfactory.
Since the Smagorinsky model has been initially developed for the convective term,
we will test it first. It is expressed as follows:
29
Cmi ≃ −2
∂Csρ ∆2|S|Sij
∂xj
, (38)
where Cs is the Smagorinsky coefficient, Sij = 0.5(∂ui/∂xj +∂uj/∂xi) is the resolved
strain rate tensor and |S| =√
2SijSij. Classically the Smagorinsky coefficient can
be directly prescribed at a constant value (Cs ≃ 0.01), or can be determined by
dynamic procedure ([Germano, 1992; Lilly, 1992]). Both methods have been tested
in this study. As presented with the LES transport equations (24; 25), ρ is not a
transported quantity. It is estimated (eq. 23) from the phase function evolution.
In figure 11, the modeled contribution of the dynamic procedure has been plotted
against the exact contribution for the SFS filter size. Both injections (fig. 11-(a-b))
and atomized (fig. 11-(c-d)) areas have been observed to evaluate the capability of
the model for various liquid densities, as for spanwise (fig. 11-a-c) and streamwise
(fig. 11-b-d) directions.
From a qualitative point of view, these scatter plots show the difficulty in finding
a clear correlation between the exact and the modeled contributions, whatever the
observed direction. In all tested cases of this model, the scatterplot is strongly
dispersed. This dispersion is worse when the filter size is large (not shown).
It appears that mixed models [Boivin et al., 2000; Toutant et al., 2008], which
include a scale similarity assumption, have a much better behavior as far as interfacial
two-phase flows are concerned. As mentioned and shown before, it appears that eddy
viscosity models are not adapted to model turbulent interfacial flows. These models
have been developed in the framework of one-phase turbulent flows without any
local discontinuities. The presence of an interface jump induces subgrid fluctuations,
which are not directly correlated to the turbulent motion, and leads to a failure of
the model. However, scale similarity models do not consider any local properties
30
of the turbulence. Indeed, they are based on the projection of resolved fluctuations
onto an unresolved grid.
Consequently, scale similarity assumptions are tested in the subsequent research.
The closures of the prevalent subgrid terms are then written:
• Convective term contribution closure:
Cmi = Cc
(∂ ρ ui uj − ρ ui uj
∂xj
). (39)
• Temporal term contribution closure:
Tmi = Ct
(∂ρ ui − ρ ui
∂t
). (40)
• Convective term of the phase function equation contribution:
Πm = Cχ
( ui
∂χ
∂xi− ui
∂χ
∂xi
), (41)
where (•) is a test filtering operator, such as ∆/∆ = 2.
One of the main drawbacks of this formulation is the apparition of a projection
coefficient between the resolved and unresolved perturbations. Contrary to usual
eddy viscosity dynamic models, there is no possible dynamic estimation of C. Thanks
to the DNS field, an estimation of this parameter is proposed in the following. Its
dependence on the filter size but also on the interface properties (continuous or
atomized) is scrutinized.
All results of the scale similarity closures have been plotted in figures 12, 13 and 14
for the convective and temporal terms of the momentum equation and the convective
31
term present in the evolution equation of the interface, respectively. Both injection
(a-b) and atomized (c-d) areas have been observed to evaluate the capability of the
model for any liquid configurations. It appears that, for all the subgrid terms, the
scale similarity assumptions offer a good agreement. All scatter plots between the
effective subgrid terms and their closures present a good correlation, contrary to the
Smagorinsky model: figure 12 has to be compared with figure 11. Similar conclusions
arise for the LFS filter size (not shown).
As aforementioned, the main drawback of the scale similarity model concerns the
estimation of the constant. It is, indeed, not possible to use a dynamical estimation
of this constant. Thanks to the DNS data, the values of the various constants of the
model as a function of the filter size were estimated. As an example, the results for
the convective term coefficient, Cc are shown in figure 15. The ratio of the filter size
to the DNS grid has been tested up to 20, although it is far from any physical value
we would effectively use. It is possible to see that, in the range of interest (between
5 and 13∆), the relation between Cc and ∆ is quasi-linear. Another important point
to note is that the values of the coefficients strongly depend on the position in the
jet.
While the two coefficients in the spanwise directions have very similar behaviors
in zone 1 to 3, the streamwise direction, which constitutes a particular direction in
this flow, is singular, especially in zone 1.
Therefore, the scale similarity model seems to be the optimal choice to carry
out LES of two-phase flows with the presence of an interface. Some values of the
various subgrid constants have been proposed, but depending on the configuration
and the atomization level, they may need some adjustment. An ideal possibility
would be to develop a model using a dynamic procedure to determine its coefficients.
However, the tests that have been performed up to now with a dynamic model are
32
not satisfactory. In fact, the presence of the interface generates spurious subgrid
information.
Moreover, it can be noted that the simple observation of the preceding scatter
plots allows, from a qualitative point of view, the different closures to be discrimi-
nated but does not state their quality. Indeed, even if the scale similarity assumption
shows better behavior than the eddy viscosity closure, it can be observed in figures
12, 13 and 14, that the results do not perfectly fit the exact contributions obtained
directly from the DNS. This is illustrated in figure 16, which shows an instantaneous
comparison between the exact streamwise contribution of the convective subgrid
term (a) and its scale similarity counterpart (b) in spanwise-spanwise cut in zone 1.
From a qualitative point of view, this closure shows good behavior compared to the
exact contribution. Still, due to the model error, point to point discrepancy can be
observed by comparing exact and modeled contributions along the y profile in figure
16-(c).
To advance further, a characterization of the statistical dispersion of the error
committed by the tested models is proposed in the following sub-section.
3.2. Statistical dispersion of the models
To characterize the statistical dispersion of the error committed by the various
models we propose to use an operator, denoted σ, on the results. An optimistic
estimation would be to suppose a certain compensation of the error. Here we opt
for the pessimistic evaluation, assuming an accumulation of the error. Thus, for our
purposes, we use the quadratic mean which permits to sum the error on the statistical
sample. Denoting DA, vector of sample data, and n, the number of samples, σ is
written as:
33
σ(A) =
√√√√ 1
n
n∑
k=1
(DAk )2 . (42)
Using this last operator, we can define a statistical error, ε, in order to characterize
the dispersion between the exact subgrid terms and the modeled ones. For instance
we can use:
ε = σ(Λ′
i − Λmi
′) , (43)
where Λ′
i and Λmi
′ denotes the exact and the modeled contributions of a given subgrid
term (e.g. Ci and Cmi ) :
Λ′
i =∂τij
∂xj
, (44)
Λmi
′ =∂τm
ij
∂xj
. (45)
We are thus seeking the model that minimises the dispersion error, ε. Relation
number 43) gives an absolute error, but it does not inform about the relative pro-
portions between exact and modeled term. We then propose to define three relative
errors, considering the fact that each subgrid term contribution Λ′
i, results from the
total unclosed subgrid correlations Λci, subtracted by the resolved part Λri
:
Λ′
i = Λci− Λri
with , (46)
Λci=
∂αiβj
∂xj, (47)
Λri=
∂αi βj
∂xj, (48)
where Λi can be written in the following form:
34
Λ′
i =∂τij
∂xj
=∂αiβj − αi βj
∂xj
. (49)
In this last relation α and β may represent any property of the flow. All the terms
of decomposition in equation 46 are summed up in table 2.
Starting from this definition, we can obtain several relative errors, denoted εsgs,
εerr and εrap :
• εsgs = σ(Λ′
i)/σ(Λci) :
Strictly speaking, εsgs is not an error, but indicates the proportion of the sub-
grid contribution Λ′
i compared to the total unclosed term Λci. It is expected
that this proportion increases with filter size, ∆.
• εerr = σ(Λ′
i − Λmi
′)/σ(Λci) :
εerr represents the proportion of the error made by the model compared to the
subgrid correlation Λci. Thus, for an ideal model this error should tend to 0.
• εrap = εerr/εsgs = σ(Λ′
i − Λmi
′)/σ(Λ′
i)
εrap gives the ratio between the two last parameters. A value under 100%
indicates that the model has a good behavior. On the contrary, a value over
100% indicates that the model overestimates, in a global sense, the subgrid
contribution.
3.2.1. Dispersion results
The multiple errors which have been obtained using the operator σ are shown
in figures 17, 18 and 19 for convective, temporal and interfacial contributions, re-
spectively. Note the use of the following common symbols notation for each tested
contribution, direction and dispersion errors:
35
• The size of the symbol is related to the filter length (i.e. smaller SFS, bigger
LFS)
• Empty and filled symbols are related to the streamwise and spanwise con-
tribution direction, respectively (except for the interfacial term, which is not
directional).
• εsgs is always represented with diamond symbols ( ld ).
Firstly, we are interested in the dispersion errors made by the various models of
convective term contribution ( figure 17).
The largest is the filter length, the highest εsgs is (fig. 17-(a)), and this for both
streamwise and spanwise directions. For example, in the streamwise direction in
zone 3, εsgs start from about 10% with SFS and up to about 20% with LFS. This
remark is also true for the temporal and convective contribution. It can be observed
in figure 18-(a) and 19-(a). This result was expected, since there is more unresolved
subgrid scales as the filter length increases.
In the streamwise direction, the convective SGS contribution is smaller in zone
3 than in zone 1 (fig. 17-(a)). Surprisingly, the opposite behavior can be observed
for the spanwise direction contributions. Since we study the divergence of convective
subgrid tensor in the i th direction, Ci = ∂τρuiuj/∂xj , this behavior could be due to
the isotropization of the velocity field in zone 3 at the small scales. Nevertheless,
there is no clear explanation of this fact and more advanced study is required to
advance the understanding of this troublesome point.
Scrutinizing εerr and εrap, in figures 17-(b) and 17-(c), reveals the fact that the
Smagorinsky model overestimates the subgrid contribution, for both constant and
dynamic procedure.
36
Nevertheless, this model leads to lower errors in zone 3 than in zone 1 (globally,
εrap > 100% in zone 1 and < 100% in zone 3). These facts can be partially explained
by the homogenization of the velocity field in zone 3. However, unlike in zone 1, we
found an inertial range in the kinetic energy spectrum in zone 3 that is absent in zone
1 (c.f. figure 6). This fact can explain the bad results of the Smagorinsky procedure
in zone 1. Nevertheless, the presence of an inertial range is implicitly considered in
the construction of the Smagorinsky model. However, we notice that in zone 3 the
dynamic procedure leads to slightly better results than the constant determination
of Cs.
Finally, concerning the scale similarity closure model, figures 17-(b) and 17-(c)
show that it offers better results than Smagorinsky models, in all considered cases.
Indeed, εrap is systematically under 100%. Moreover, the quality of this closure
is less affected by the considered direction or zone than the Smagorinsky model.
For example in the SFS case, εrap is close to 73% in zone 1 and 3 for streamwise
contributions. For these last terms, in the LFS case, we found 79% and 85% for zone
1 and 3, respectively. Therefore, the filter size does not affect the model accuracy.
Thus, the scale similarity model seems to be much more versatile and it permits
to obtain better estimations of subgrid contributions than the Smagorinsky models.
However, we have to keep in mind that the scale similarity constant is optimised for
each case thanks to the DNS.
Starting from this conclusion, we now consider errors obtained by the scale sim-
ilarity model for the temporal contribution Ti, in figure 18. The behavior of εsgs for
this term is fully similar to the one obtained for convective contribution.
The scale similarity closure gives an error εrap under 100% in all tested cases. For
all directions, this error is about 70% for SFS cases and about 80% in LFS cases.
37
The last tested term is the interfacial contribution. Note that there is no direction
aspect since we observe the dispersion errors for Π = ui∂χ/∂xi − ui ∂χ/∂xi. The
results obtained for this term are plotted in figure 19. Firstly, the study of εsgs reveals
that the proportion of subgrid contributions to close becomes more important with
increasing filter size, in the same manner as the other two contributions. We can also
notice, the fact that this proportion is bigger in zone 1 than in zone 3. Again, the
scale similarity model permits to obtain dispersion error, εerr and εrap, values under
100% for each tested zone.
The study of the dispersion errors obtained in this section clearly shows that the
scale similarity model is more able to reproduce the convective subgrid contribution
of transport equations than the usual Smagorinsky model, even with dynamic proce-
dure. Once the constant Cc, Ct and Cχ are correctly determined, the application of
this model on temporal and interfacial terms also shows good behaviors. From this
point of view, the conclusions here are in full agreement with other a priori studies
([Boivin et al., 2000; Toutant et al., 2008]) in different two-phase flow configurations.
These studies suggest that the scale similarity model or mixed model, which include
this assumption, lead to better estimations of subgrid contributions than the classical
eddy viscosity models. We thus reach the same conclusion for a liquid/gas flow in
an atomization regime.
Nevertheless, we can recall the fact that a priori testing, in the way it has been
led, implies some drawbacks. On the one hand, the validity of the results obtained
here can only be assumed for the tested Reynolds number, which is relatively low.
The quality of the subgrid closure has to be confirmed for higher Reynolds or Weber
numbers. On the other hand, the high computational of DNS did not permit to
take into account the largest time scale interactions between the subgrid terms.
Consequently, a possible accumulation of errors made by a closure on long time
38
simulations cannot be detected with this sort of a priori test.
Assuming an accumulation of the errors, we notice that the dispersion errors
found even with the scale similarity model are somewhat higher (more than 50%
for εsgs). Moreover, the gain obtained when applying the scale similarity model by
comparison to directly neglect the subgrid terms is quite low. In fact, the mean error
εrap range lies between 60% and 80%. Thus, it could be interesting to test other
subgrid models, like the mixed model [Bardina et al., 1980], in order to know if it is
possible to obtain better results.
4. Conclusions
In this paper, the direct numerical simulation of a fully atomizing jet of liquid
has been carried out to evaluate, from an a priori point of view, the various subgrid
properties from the liquid destabilization area, close to the injector, down to the fully
atomized jet. Simulations have been performed thanks to the coupling of level-set
and VOF formulations.
An a priori analysis of the subgrid terms has been carried out to evaluate the order
of magnitude of the contributions of these terms to the flow evolution. It is the first
time that this kind of analysis is carried out on such a complex configuration. The
results show that, wherever we are positioned in the atomizing jet, the hierarchy
between the subgrid terms of the momentum equation remains the same with a
prevalence of the temporal and the convective terms. This observation confirms
previous works carried out on simpler configurations. The evolution equation of the
filtered composition field has been scrutinized as well. It appears that the interface
evolution is not mainly controlled by large-scale convection phenomena. Thus, it
is necessary to take into account the impact of the small scales on the interface
propagation. Furthermore, at the interface level close to the injection, the order of
39
magnitude of the subgrid correlations increases strongly. Knowing that the initial
liquid destabilization process could appear at a subgrid level, it may be dangerous
to neglect them. Therefore a closure has been proposed and evaluated as well. The
choice has been made to use a scale similarity assumption to close all the subgrid
unresolved terms. Classical dynamic and static Smagorinsky formulations have been
tested as well, but the results were not satisfactory at all because of the presence of
the interface that leads to strong gradients, which are not related to the turbulent
motion. The scale similarity assumption gave better results as far as the DNS of
an atomizing jet is concerned. This means that the geometry dimensions remain
small with a liquid Reynolds number equal to 5800. One of the main encouraging
conclusions is the fact that the models are able to capture the subgrid terms wherever
we are in the jet: in the atomization area or far from the injector.
To go further, a priori tests have been extended by the analysis of the statistical
dispersion of the error made by the various models. The choice has been made to
consider a possible accumulation of the error. The clear benefit to using the scale sim-
ilarity model by comparison with the eddy viscosity model has been demonstrated.
However, even if the scale similarity model brings systematically an improvement,
errors due to the model remain. This implies that more efforts should be necessary
to improve subgrid atomization modeling for LES of atomization.
Very encouraging results have been observed and the next stage is now to carry
out effective LES with a posteriori validations and the development of a Eulerian
procedure to describe the dispersion of the droplets reaching a subgrid size.
Acknowledgement
This work was granted access to the HPC resources of IDRIS under the allocation
2010-x2010026153 made by GENCI (Grand Equipement National de Calcul Intensif).
40
The authors also acknowledge the computational support of CRIHAN (Centre de
Ressources Informatiques de Haute-Normandie) and Renault S.A. for their financial
support.
41
References
Bardina, J., Ferziger, J., Reynolds, W., 1980. Improved subgrid-scale models for
large-eddy simulation. AIAA paper-80.
Baumgarten, C., 2006. Mixture formation in Internal Combustion Engine. Springer.
Berlemont, A., Tanguy, S., 2005. Development of a level set method for interface
tracking : applications to droplet collisions. Int. J. Multiphase Flows 31, 1015–
1035.
Bianchi, G., Minelli, F., Scardovelli, R., Zaleski, S., 2007. 3d large scale simulation
of the high-speed liquid jet atomization. SAE international2007-01-0244.
Boivin, M., Simonin, O., Squires, K. D., 2000. On the prediction of gas–solid flows
with two-way coupling using large eddy simulation. Physics of Fluids 12 (8), 2080–
2090.
Brackbill, J., Khote, D., Zemach, C., 1992. A continuum method for modelling
surface tension. J. Comput. Phys. 100, 335–354.
Buonfiglioli, M., Mendonca, F., 2005. LES-VOF simulation of primary diesel spray
break-up with synthetic inlet perturbations. In: ILASS america.
Chesnel, J., Menard, T., Reveillon, J., Berlemont, A., Demoulin, F.-X., 2010. A les
simulation of atomisation. In: ASME, Montreal. No. FEDSM-ICNMM2010-30664.
Chesnel, J., Reveillon, J., Demoulin, F. X., Menard, T., 2007. Subgrid modeling of
liquid atomization. In: 6th International Conference on Multiphase Flow.
Christensen, E. D., 2006. Large eddy simulation of spilling and plunging breakers.
Coastal Engineering 53 (5-6), 463 – 485.
42
De Villier, E., Gosman, A., Weller, H., 2004. Large eddy simulation of primary diesel
spray atomisation. SAE international 113 (3), 193–206.
Deen, N., Solberg, T., Hjertager, B., 2001. Large eddy simulation of the gas-liquid
flow in a square cross-sectionned bubble column. Chemical Engineering science 5,
6341–6349.
Delhaye, J. M., 1974. Jump conditions and entropy sources in two-phase systems.
local instant formulation. International Journal of Multiphase Flow 1 (3), 395 –
409.
Demoulin, F., Beau, P., Blokkeel, G., Mura, A., Borghi, R., 2007. A new model for
turbulent flows with large density fluctuations: application to liquid atomization.
Atom. and Sprays 17, 315–345.
Drew, D., 1983. Mathematical modeling of two phase flow. Ann. Rev. Fluid Mech.
15, 261 – 291.
Faeth, G. M., 1991. Structure and atomization properties of dense turbulent sprays.
In: Symposium (International) on Combustion, 23rd, Orleans, France, July 22-
27, 1990, Proceedings (A92-16529 04-25). Pittsburgh, PA, Combustion Institute,
1991, p. 1345-1352. pp. 1345–1352.
Fedkiw, R. P., Aslam, T., Merriman, B., Osher, S., 1999. A non-oscillatory eulerian
approach to interfaces in multimaterial flows (the ghost fluid method). Journal of
Computational Physics 152 (2), 457 – 492.
Germano, M., 1992. Turbulence: the filtering approach. J. Fluid Mech. 238, 325–336.
Germano, M., Piomelli, U., Moin, P., Cabot, W., 1991. A dynamic subgride-scale
eddy viscosity model. Phy. Fluids 7(3), 1760–1765.
43
Herrmann, M., Gorokhovski, M., 2008. An outline of a les subgrid model for liq-
uid/gas phase interface dynamics. In: Proceedings of the 2008 CTR Summer Pro-
gram, Center for Turbulence Research, Stanford University, Stanford, CA. pp.
171–181.
Hieu, P., Katsutoshi, T., Ca, V., 2004. Numerical simulation of breaking waves using
a two phase flow model. appl. Math. Model. 28, 983–1005.
Hirt, C., Nichols, B., 1981. Volume of fluid (vof) - method for the dynamics of free
boundaries. J. Comput. Phys. 39, 201–225.
Kang, M., Liu, X., Fedkiw, R., 2000. Boundary condition capturing method for
multiphase imcompressible flows. J. Sci. Comput. 15, 323–360.
Kataoka, I., 1986. Local instant formulation of two-phase flow. International Journal
of Multiphase Flow 12 (5), 745 – 758.
Klein, M., Sadiki, A., Janicka, J., 2003. A digital filter based generation of inflow
data for spatially developing direct numerical or large eddy simulations. Journal
of Computational Physics 186 (2), 652 – 665.
Labourasse, E., Lacanette, D., Toutant, A., Lubin, P., Vincent, S., Lebaigue, O.,
Caltagirone, J.-P., Sagaut, P., 2007. Towards large eddy simulation of isothermal
two-phase flows: Governing equations and a priori tests. International Journal of
Multiphase Flow 33 (1), 1 – 39.
Lakehal, D., Meier, M., Fulgosi, M., 2002. Interface tracking towards the direct
simulation of heat and mass transfer in multiphase flows. International Journal of
Heat and Fluid Flow 23 (3), 242 – 257.
44
Larocque, J., Vincent, S., Lacanette, D., Lubin, P., Caltagirone, J.-P., 2010. Para-
metric study of les subgrid terms in a turbulent phase separation flow. International
Journal of Heat and Fluid Flow In Press, Corrected Proof, –.
Law, C., 2006. Combustion Physics. Cambridge University Press.
Lebas, R., Menard, T., Beau, P., Berlemont, A., Demoulin, F., 2009. Numerical
simulation of primary break-up and atomization: DNS and modelling study. In-
ternational Journal of Multiphase Flow 35 (3), 247 – 260.
Lee, S., Lele, K., Moin, P., 1991. Numerical simulations of spatially evolving com-
pressible turbulence. Center for Turbulence Research, Annual Research Briefs,
Stanford 126.
Lilly, D., 1992. A proposed modification of the germano subgridscale closure method.
Phys. Fluid A 4, 633–635.
Liovic, P., Lakehal, D., 2007. Interface-turbulence interactions in large-scale bubbling
processes. International Journal of Heat and Fluid Flow 28 (1), 127 – 144, the
International Conference on Heat Transfer and Fluid Flow in Microscale (HTFFM-
05), The International Conference on Heat Transfer and Fluid Flow in Microscale.
Lubin, P., Vincent, S., Abadie, S., Caltagirone, J.-P., 2006. Three-dimensional large
eddy simulation of air entrainment under plunging breaking waves. Coastal Engi-
neering 53 (8), 631 – 655.
Menard, T., Tanguy, S., Berlemont, A., 2007. Coupling level set/vof/ghost fluid
methods: Validation and application to 3d simulation of the primary break-up of
a liquid jet. International Journal of Multiphase Flow 33, 510–524.
45
Nicoud, F., Ducros, F., 1999. Subgrid-scale stress modelling based on the square of
the velocity gradient tensor. Flow, Turbulence and Combustion 62, 183–200(18).
Orszag, S., Patterson, G., Jan 1972. Numerical simulation of three-dimensional ho-
mogeneous isotropic turbulence. Phys. Rev. Lett. 28 (2), 76–79.
Osher, S., Sethian, J., 1988. Fronts propagating with curvature-dependent speed: Al-
gorithms based on hamilton-jacobi formulations. Journal of Computational Physics
79 (1), 12 – 49.
Reitz, R. D., 1978. Atomization and other break-up regimes of a liquid jet. Ph.D.
thesis, Princeton University.
Rogallo, R. S., 1981. Numerical experiments in homogeneous turbulence. Tech. rep.,
NASA, technical Memorandum No. 81315.
Sagaut, P., 2003. Large Eddy Simulation for Imcompressible Flows. Springer Verlag.
Sagaut, P., Germano, M., 2005. On the filtering paradigm for les of flows with dis-
continuities. Journal of Turbulence 6, N23.
Sagaut, P., Montreuil, E., Labb, O., 1999. Assessment of some self-adaptive sgs
models for wall bounded flows. Aerospace Science and Technology 3 (6), 335 –
344.
Sirignano, W. A., 1983. Fuel droplet vaporization and spray combustion theory.
Progress in Energy and Combustion Science 9, 291–322.
Smagorinsky, J., 1963. General experiments with the primitives equations. Mon.
Weather Rev. 63, 99–164.
46
Sussman, M., Puckett, E. G., 2000. A coupled level set and volume-of-fluid method
for computing 3d and axisymmetric incompressible two-phase flows. Journal of
Computational Physics 162 (2), 301 – 337.
Tanguy, S., T., M., Berlemont, A., 2007. A level set method for vaporizing two-phase
flows. Journal of Computational Physics 221 (2), 837 – 853.
Toutant, A., Labourasse, E., Lebaigue, O., Simonin, O., 2008. DNS of the interaction
between a deformable buoyant bubble and a spatially decaying turbulence: A priori
tests for les two-phase flow modelling. Computers and Fluids 37 (7), 877 – 886,
special Issue of the ”Turbulence and Interaction-TI2006” Conference.
Vallet, A., Burluka, A., Borghi, R., 2001. Developpement of an eulerian model for
the atomization of a liquid jet. Atom. and Sprays 11, 619–642.
van der Pijl, S. P., Segal, A., Vuik, C., Wesseling, P., 2005. A mass-conserving level-
set method for modelling of multi-phase flows. International Journal for Numerical
Methods in Fluids 47, 339–361.
Vincent, S., Larocque, J., Lacanette, D., Toutant, A., Lubin, P., Sagaut, P., 2008.
Numerical simulation of phase separation and a priori two-phase les filtering.
Computers and Fluids 37 (7), 898 – 906, special Issue of the ”Turbulence and
Interaction-TI2006” Conference.
Wang, P., Frhlich, J., Michelassi, V., Rodi, W., 2008. Large-eddy simulation of
variable-density turbulent axisymmetric jets. International Journal of Heat and
Fluid Flow 29 (3), 654 – 664, the Fifth International Symposium on Turbulence
and Shear Flow Phenomena (TSFP5).
47
Yue, W., Lin, C.-L., Patel, V. C., 2005. Large eddy simulation of turbulent open-
channel flow with free surface simulated by level set method. Physics of Fluids
17 (2), 025108.
Injector holex2
x3
x1
Ld ≃ 24D0
Zone 1 Zone 2 Zone 3
Liquid jet contour
D0
Figure 1: Computational domain.
48
Figure 2: DNS of a liquid jet atomization
49
Parameter Symbol/Unit Value
Gas density ρg(kg.m−3) 25
Liquid density ρl(kg.m−3) 696
Gas viscosity µg(kg.m−1s−1) 1.x10−5
Liquid viscosity µl(kg.m−1s−1) 1.18x10−3
Surface tension coefficient σ(N.m−1) 0.06
Injection Diameter D0(µm) 100
Mean flow rate velocity Ud(m.s−1) 79
Liquid Reynolds Rel 4659
Liquid Weber Wel 7239
Turbulence Intensity u′u′/U2 0.05
Turbulent scale Lt(m) 0.1D0
Table 1: Physical parameters
50
5 10 15 20−1
01
0
0.5
1
y/D
0
x/D0
(a) DNS field
5 10 15 20−1
01
0
0.5
1
y/D
0
x/D0
(b) ∆ = 5∆.
5 10 15 20−1
01
00.51
y/D
0
x/D0
(c) ∆ = 13∆.
Figure 3: Impact of filtering on the phase variable, streamwise direction (x−z) cuts. The
white line represents the interface position. (a) : DNS field, (b) : SFS ∆ = 5∆, (c) : LFS
∆ = 13∆.
51
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
100
200
300
400
500
600
z/D0
y/D
0
(a) ρ DNS, zone 1.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
100
200
300
400
500
600
z/D0
y/D
0
(b) ρ DNS, zone 2.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
100
200
300
400
500
600
z/D0
y/D
0
(c) ρ DNS, zone 3.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
100
200
300
400
500
600
z/D0
y/D
0
(d) ρ(χ), zone 1.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
100
200
300
400
500
600
z/D0
y/D
0
(e) ρ(χ), zone 2.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
100
200
300
400
500
600
z/D0
y/D
0
(f) ρ(χ), zone 3.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
100
200
300
400
500
600
z/D0
y/D
0
(g) ρ(ε), zone 1.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
100
200
300
400
500
600
z/D0
y/D
0
(h) ρ(ε), zone 2.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
100
200
300
400
500
600
z/D0
y/D
0
(i) ρ(ε), zone 3.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
50
100
150
200
250
300
z/D0
y/D
0
(j) ρ(ε) subgrid, zone 1.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
50
100
150
200
250
300
z/D0
y/D
0
(k) ρ(ε) subgrid, zone 2.
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
50
100
150
200
250
300
z/D0
y/D
0
(l) ρ(ε) subgrid, zone 3.
Figure 4: Densities ρ and ρ on spanwise cuts xy for zones 1, 2 and 3.
52
0 1 2 3 4 50
0.01
0.02
0.03
0.04
0.05
0.06
0
0=1.5
0=12.8
0=22.6
Ener
gy〈k
e〉u
−2
0
r/r
x/D
x/D
x/D
(a) Turbulent kinetic energy, 〈ke〉
0 1 2 3 4 50
1
2
3
4
5
6
0
0=1.5
0=12.8
0=22.6
Dis
sipat
ion〈ε〉d
−2
0t−
30
r/r
x/D
x/D
x/D
(b) Dissipation, 〈ε〉
Figure 5: Turbulence properties evolution along the streamwise direction. Three positions
: 1.5, 12.8 and 22.6D0.
0,01 0,1 1
0,0001
0,01
1
E/Emax
k/kmax
k13∆ k5∆
Figure 6: Kinetic energy spectrum. Dot-dashed straight line : k−5/3, line with symbol :
injection area (0 - 4 D0), dashed line : central area (8 - 12 D0), solid line : atomized area
(20 - 24 D0).
53
SGS term Λ′i
τρuiujC ′
i = ∂(ρuiuj − ρ ui uj)/∂xj
τρuiT ′
i = ∂(ρui − ρ ui)/∂t
τχ Π′ = ui∂χ/∂xi − ui ∂χ/∂xi
SGS term Λci
τρuiujCci
= ∂ρuiuj/∂xj
τρuiTci
= ∂ρui/∂t
τχ Πc = ui∂χ/∂xi
SGS term Λri
τρuiujCri
= ∂ρ ui uj/∂xj
τρuiTri
= ∂ρ ui/∂t
τχ Πr = ui ∂χ/∂xi
Table 2: Expressions of the resolved/unresolved decomposition terms (eq. 46) of the
convective, temporal and interfacial contributions.
54
0 1 2 3 4 50
0.2
0.4
0.6
0.8
1
〈k/k
〉
r/r0
(a) Position 1.5D0, zone 1.
0 1 2 3 4 50
0.2
0.4
0.6
0.8
1〈k
/k〉
r/r0
(b) 18.5D0, zone 3.
Figure 7: Ratio of the resolved kinetic energy k with the total kinetic energy k, radial
mean. Solid lines : ∆ = 5∆, dashed : ∆ = 13∆.
55
0 1 2 3 4 510
−10
10−5
100
105
0
r/r
Con
trib
uti
on(k
g.m
−2.s
−2)
temporalconvectivediffusiveinterfacial
(a) Streamwise momentum ρu1, position : 5
D0
0 1 2 3 4 510
−10
10−5
100
105
0
r/r
Con
trib
uti
on(k
g.m
−2.s
−2)
temporalconvectivediffusiveinterfacial
(b) Streamwise momentum ρu1, position 20
D0
0 1 2 3 4 510
−10
10−5
100
105
0
r/r
Con
trib
uti
on(k
g.m
−2.s
−2)
temporalconvectivediffusiveinterfacial
(c) Spanwise momentum ρu3 : 5 D0
0 1 2 3 4 510
−8
10−6
10−4
10−2
100
102
0
r/r
Con
trib
uti
on(k
g.m
−2.s
−2)
temporalconvectivediffusiveinterfacial
(d) Spanwise momentum ρu3 : 20 D0
Figure 8: Subgrid contributions for momentum equation. ∆ = 5∆. log scale.
56
Liquid
Gas
(a) τρu1u2, xy cut in zone 1.
Liquid
Gas
(b) τρu1u2, yz cut.
Liquid
Gas
(c) τρu2u2, xy cut in zone 1.
Liquid
Gas
(d) τρu2u2, yz cut.
Figure 9: Thin lines : momentum subgrid contribution estimation (plain lines : τ > 0,
dashed lines τ < 0), LFS case. Thick line : smoothed interface contour. Minimum and
maximum values for τρu1u2are −1.6×105 and 2.4×105kg.m−2.s−2, respectively. Minimum
and maximum values for τρu2u2are 2 × 103 and 2 × 105kg.m−2.s−2, respectively.
57
0 1 2 3
0,01
0,1
1
r/r0
Con
trib
uti
ons
(×10
4s−
1)
(a) Zone 1.
0 1 2 3
0,01
0,1
1
r/r0
(b) Zone 3.
0 1 2 30,001
0,01
0,1
1
r/r0
Con
trib
uti
ons
(×10
4s−
1)
(c) Zone 1.
0 1 2 30,001
0,01
0,1
1
r/r0
(d) Zone 3.
Figure 10: Resolved (solid lines) and subgrid (dashed lines) contributions to the phase
function evolution equation. Zone 1 (a) and 3 (b) : ∆ = 5∆. Zone 1 (c) and 3 (d) :
∆ = 13∆
58
−1e+00 −5e−01 0 5e−01 1e+00−1e+00
−5e−01
0
5e−01
1e+00
14.722.0150.1
−2∂
Csρ∆|S|S
2j/∂
xj
∂(ρu2uj − ρ u2 uj)/∂xj
εsgs =εerr =εrap = %
%%
(a) Spanwise direction, zone 1.
−3e+00 −2e+00 0 2e+00 3e+00−3e+00
−2e+00
0
2e+00
3e+00
27.130.0110.5
−2∂
Csρ∆|S|S
1j/∂
xj
∂(ρu1uj − ρ u1 uj)/∂xj
εsgs =εerr =εrap = %
%%
(b) Streamwise direction, zone 1.
−2e+00 −1e+00 0 1e+00 2e+00−2e+00
−1e+00
0
1e+00
2e+00
35.230.486.5
−2∂
Csρ∆|S|S
2j/∂
xj
∂(ρu2uj − ρ u2 uj)/∂xj
εsgs =εerr =εrap = %
%%
(c) Spanwise direction, zone 3.
−3e+00 −2e+00 0 2e+00 3e+00−3e+00
−2e+00
0
2e+00
3e+00
9.48.489.2
−2∂
Csρ∆|S|S
1j/∂
xj
∂(ρu1uj − ρ u1 uj)/∂xj
εsgs =εerr =εrap = %
%%
(d) Streamwise direction, zone 3.
Figure 11: Dispersion diagrams for the convective term with dynamic Smagorinsky pro-
cedure. ∆ = 5∆x.
59
−6e−01 −3e−01 0 3e−01 6e−01−6e−01
−3e−01
0
3e−01
6e−01
13.18.665.4
∂(
ρu
2u
j−
ρu
2u
j)/
∂x
j
∂(ρu2uj − ρ u2 uj)/∂xj
εsgs =εerr =εrap = %
%%
(a) Spanwise direction, zone 1.
−4e+00 −2e+00 0 2e+00 4e+00−4e+00
−2e+00
0
2e+00
4e+00
29.621.672.9
∂(
ρu
1u
j−
ρu
1u
j)/
∂x
j
∂(ρu1uj − ρ u1 uj)/∂xj
εsgs =εerr =εrap = %
%%
(b) Streamwise, zone 1.
−2e+00 −1e+00 0 1e+00 2e+00−2e+00
−1e+00
0
1e+00
2e+00
34.924.570.2
∂(
ρu
2u
j−
ρu
2u
j)/
∂x
j
∂(ρu2uj − ρ u2 uj)/∂xj
εsgs =εerr =εrap = %
%%
(c) Spanwise direction, zone 3.
−3e+00 −2e+00 0 2e+00 3e+00−3e+00
−2e+00
0
2e+00
3e+00
10.67.772.7
∂(
ρu
1u
j−
ρu
1u
j)/
∂x
j
∂(ρu1uj − ρ u1 uj)/∂xj
εsgs =εerr =εrap = %
%%
(d) Streamwise direction, zone 3.
Figure 12: Dispersion diagrams for the convective term with scale similarity assumption.
∆ = 5∆x.
60
−7e−01 −3e−01 0 3e−01 7e−01−7e−01
−3e−01
0
3e−01
7e−01
13.09.371.5
∂(ρ
u2−
ρu
2)/
∂t
∂(ρu2 − ρ u2)/∂t
εsgs =εerr =εrap = %
%%
(a) Spanwise direction, zone 1.
−2e+00 −1e+00 0 1e+00 2e+00−2e+00
−1e+00
0
1e+00
2e+00
16.911.567.9
∂(ρ
u1−
ρu
1)/
∂t
∂(ρu1 − ρ u1)/∂t
εsgs =εerr =εrap = %
%%
(b) Streamwise direction, zone 1.
−3e+00 −2e+00 0 2e+00 3e+00−3e+00
−2e+00
0
2e+00
3e+00
38.827.170.0
∂(ρ
u2−
ρu
2)/
∂t
∂(ρu2 − ρ u2)/∂t
εsgs =εerr =εrap = %
%%
(c) Spanwise direction, zone 3.
−3e+00 −2e+00 0 2e+00 3e+00−3e+00
−2e+00
0
2e+00
3e+00
5.94.372.3
∂(ρ
u1−
ρu
1)/
∂t
∂(ρu1 − ρ u1)/∂t
εsgs =εerr =εrap = %
%%
(d) Streamwise direction, zone 3.
Figure 13: Dispersion diagrams for the temporal term with scale similarity assumption.
∆ = 5∆x.
61
−2e−06 −1e−06 0 1e−06 2e−06−2e−06
−1e−06
0
1e−06
2e−06
13.29.672.7
ui
∂χ
∂x
i−
ui
∂χ
∂x
i
ui∂χ∂xi
− ui∂χ∂xi
εsgs =εerr =εrap = %
%%
(a) ∆ = 5∆x, zone 1.
−2e−06 −1e−06 0 1e−06 2e−06−2e−06
−1e−06
0
1e−06
2e−06
33.828.584.6
u
i∂χ
∂x
i−
ui
∂χ
∂x
i
ui∂χ∂xi
− ui∂χ∂xi
εsgs =εerr =εrap = %
%%
(b) ∆ = 13∆x, zone 1.
−3e−06 −2e−06 0 2e−06 3e−06−3e−06
−2e−06
0
2e−06
3e−06
6.85.276.5
ui
∂χ
∂x
i−
ui
∂χ
∂x
i
ui∂χ∂xi
− ui∂χ∂xi
εsgs =εerr =εrap = %
%%
(c) ∆ = 5∆x, zone 3.
−9e−07 −4e−07 0 4e−07 9e−07−9e−07
−4e−07
0
4e−07
9e−07
14.813.288.8
u
i∂χ
∂x
i−
ui
∂χ
∂x
i
ui∂χ∂xi
− ui∂χ∂xi
εsgs =εerr =εrap = %
%%
(d) ∆ = 13∆x, zone 3.
Figure 14: Dispersion diagrams for the interfacial term with scale similarity assumption.
62
0 5 10 15 202
2.5
3
3.5
4
4.5
5
5.5
Cc
∆/∆
Direction 3Direction 2Direction 1
(a) Zone 1
0 5 10 15 202
3
4
5
6
7
Cc
∆/∆
Direction 3Direction 2Direction 1
(b) Zone 3
Figure 15: Estimation of the scale similarity coefficients with respect to the subgrid filter
size.
63
-0.6 -0.4 -0.2 0 0.2 0.4 0.6-0.6
-0.4
-0.2
0
0.2
0.4
0.6
y/D0
z/D
0
(a) Actual contribution C1.
-0.6 -0.4 -0.2 0 0.2 0.4 0.6-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.80.60.40.20.01
y/D0
z/D
0
(b) Modeled contribution, Cm
1.
-1 -0.5 0 0.5 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
y/D0
Con
trib
uti
on
(c) profiles of C1 (symbol line) and Cm
1(dashed line)
along y direction.
Figure 16: (a-b) : yz cut in zone 1 of actual and modeled convective subgrid contribution
in the streamwise direction, respectively. (a) : Actual term, C1. (b) : modeled term Cm1
with scale similarity assumption. (c) : profiles of exact and modeled contribution along y
direction (location visible in dashed line on (b)). ∆ = 13∆x.
64
ε(%
)
zone 1 zone 30
20
40
60
80
100
(a) εsgs
ε(%
)
zone 1 zone 30
20
40
60
80
100
(b) εerr
ε(%
)
zone 1 zone 30
100
50
150
200
250
300
(c) εrap
Figure 17: Dispersion errors for the convective terms in zone 1 and 3. Empty and filled
symbols are for the streamwise and spanwise contributions, respectively. The scale of the
symbols is related to the filter size (i.e. smaller: SFS and biggest: LFS). ( ut ) : Smagorinsky
model; ( bc ) : Eddy viscosity model, dynamic procedure. ( rs ) : Scale similarity model.
65
ε(%
)
zone 1 zone 30
20
40
60
80
100
(a) εsgs
ε(%
)
zone 1 zone 30
20
40
60
80
100
(b) εerr, εrap
Figure 18: Dispersion errors for the temporal term in zone 1 and 3 using the scale
similarity assumption. Empty and filled symbols are for the streamwise and spanwise
contributions, respectively. The scale of the symbols is related to the filter size (i.e. smaller:
SFS and biggest: LFS). ( ut ) : εerr; ( ut ) : εrap.
66
ε(%
)
zone 1 zone 30
20
40
60
80
100
(a) εsgs
ε(%
)
zone 1 zone 30
20
40
60
80
100
(b) εerr, εrap
Figure 19: Dispersion errors for the interfacial term in zone 1 and 3. The scale of the
symbols is related to the filter size (i.e. smaller: SFS and biggest: LFS). ( u ) : εerr; ( u )
: εrap.
67